CFRNet: Cycle-Consistent Fixed-Point Training for Real-Time Blind Face Restoration on Consumer Embedded NPUs

2026-06-08 · Source: cs.CV updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems · Depth: Expert, extended

Summary

CFRNet is a 2.0 M-parameter ResNet-style neural network designed for real-time blind face restoration on consumer embedded NPUs at a $256\times 256$ resolution. It introduces Cycle-Consistent Fixed-Point Training (CCFP), which trains the network to act as a fixed-point operator, ensuring that repeated application does not alter an already restored face. CCFP incorporates progressive multi-cycle supervision, an idempotence loss, and a re-degradation cycle loss, adding no inference cost. On a 300-image test set, CFRNet achieves the best perceptual score (LPIPS 0.250 at three cycles, 31% lower than one cycle) and the best PSNR and SSIM at two cycles among deploy-compatible baselines. It runs in approximately 23 ms per cycle in INT8 on a HiSilicon Hi3402 NPU, where other "Lite" baselines fail to compile. The cycle count $k$ serves as a quality knob, with PSNR peaking at $k=2$ and LPIPS improving up to $k=3$.

Key takeaway

For Machine Learning Engineers deploying face restoration models on embedded NPUs, CFRNet offers a robust solution. You should consider its Cycle-Consistent Fixed-Point Training (CCFP) approach to achieve high perceptual quality (LPIPS 0.250 at $k=3$) and NPU compatibility. You can dynamically adjust the cycle count $k$ at inference to balance perceptual quality ($k=3$) with pixel fidelity ($k=2$) without retraining, optimizing for your specific application's latency and quality needs.

Key insights

Training a face restorer as a fixed-point operator ensures stable, iterative refinement on resource-constrained NPUs.

Principles

• Iterative inference benefits from fixed-point training.
• Perception-distortion trade-off exists within one model.
• Local component supervision improves artifact reduction.

Method

Cycle-Consistent Fixed-Point Training (CCFP) uses progressive multi-cycle supervision, an idempotence loss, and a re-degradation cycle loss to train a generator as a fixed-point operator.

In practice

• Use $k=3$ for best perceptual quality (LPIPS).
• Use $k=2$ for best pixel fidelity (PSNR/SSIM).
• Add nose region supervision for mid-face artifacts.

Topics

• Blind Face Restoration
• Neural Processing Units
• Cycle-Consistent Fixed-Point Training
• Real-time Inference
• INT8 Quantization
• Embedded AI

Best for: Computer Vision Engineer, Research Scientist, AI Scientist, Machine Learning Engineer, AI Hardware Engineer

Related on AIssential

• CFRNet: Cycle-Consistent Fixed-Point Training for Real-Time Blind Face Restoration on Consumer Embedded NPUs — CFRNet uses Cycle-Consistent Fixed-Point Training to enable high-quality, real-time blind face restoration on consumer NPUs with a tunable quality knob.
• Support-Conditioned Flow Matching Is Kernel Smoothing — Cross-attention conditioning in generative models is kernel smoothing, with predictable failure modes and learned corrections.
• Consistent-Inversion: Reverse Consistency Guidance for Structure-Preserving Visual Editing — Reverse consistency guidance corrects structural drift in diffusion-based image editing by checking trajectory reversibility.
• PVRF: All-in-one Adverse Weather Removal via Prior-modulated and Velocity-constrained Rectified Flow — PVRF unifies adverse weather removal by combining VLM-derived soft perceptions with a perception-guided rectified flow for enhanced fidelity and realism.
• Provable Offline Reinforcement Learning for Structured Cyclic MDPs — CycleFQI offers a modular, theoretically robust offline RL solution for cyclic MDPs with heterogeneous stages.
• Phase Marginalization for Patch-Grid Instability in Vision Transformers — Phase Marginalization (PM) is a training-free post-hoc method that stabilizes Vision Transformer dense predictions by accounting for patch-grid phase variations.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CV updates on arXiv.org.